Back to Gap Map Model components / Base / pretrained models

Seed-OSS

open weights / Overall score: 3.4

ByteDance Seed's open-weight model line, released August 2025 as 36B base and instruct checkpoints trained on 12T tokens. The instruct model targets long-context, reasoning and agentic use with a controllable thinking budget, and the base ships in with- and without-synthetic-data variants aimed at fine-tuning research.

ByteDance's open-weight counterpart to the closed doubao-seed line, and the openness contrast between the two is the org-page signal. Apache-2.0 weights with no released corpus or training pipeline; the tech report was still "coming soon" on the card at read time.

Openness

3 high confidence

3.0

weights: open(36B Base, Base-woSyn and Instruct, ungated on HF)
data: closed(12T-token pretraining corpus not released
code: partial(ByteDance-Seed/seed-oss is inference/usage
license: Apache-2.0(OSI)

Seed-OSS ships Apache-2.0 weights across all three SKUs, ungated, but the 12T-token training corpus is unreleased and only inference code is published. The model card states plainly that the series is released "under the Apache-2.0 license" while its technical report on the training recipe is still coming.

https://api.github.com/repos/ByteDance-Seed/seed-oss recorded 2026-09-01
`license: Apache-2.0`, 892 stars; the release repository, carrying no training pipeline — what holds `code` at partial.
https://huggingface.co/api/models/ByteDance-Seed/Seed-OSS-36B-Instruct recorded 2026-10-07
Model API for ByteDance-Seed/Seed-OSS-36B-Instruct: `gated: false`, `private: false`, cardData.license `apache-2.0`, safetensors weights listed.
https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct/raw/main/README.md recorded 2026-10-07
Card: "We release `Seed-OSS-36B-Base` (both with and without synthetic data versions) and `Seed-OSS-36B-Instruct`"; "Although trained with only 12T tokens"; "We release this series of models to the open-source community under the Apache-2.0 license."; and a "Tech Report Coming Soon" badge, with no corpus or pipeline link.

Adoption

2 high confidence

2.0

Adoption is measured as Hugging Face downloads summed across the two declared checkpoints, the instruct and base 36B releases.

https://huggingface.co/api/models/ByteDance-Seed/Seed-OSS-36B-Base recorded 2026-09-01
6,796 downloads in the trailing 30 days for ByteDance-Seed/Seed-OSS-36B-Base
https://huggingface.co/api/models/ByteDance-Seed/Seed-OSS-36B-Instruct recorded 2026-09-01
35,117 downloads in the trailing 30 days for ByteDance-Seed/Seed-OSS-36B-Instruct

Capability

4 medium confidence

4.0

The strong 30B-class open-weight tier: card benchmarks put the instruct model at or ahead of same-size open peers on math and reasoning despite a smaller training budget, which is the same "best of the ~32B open class, short of the trillion-parameter frontier" reading that puts olmo at 4. Vendor-published numbers with the tech report still pending, so confidence stays medium.

https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct/raw/main/README.md recorded 2026-10-07
Card benchmark tables with AIME24 91.7 for Seed-OSS-36B-Instruct against same-size open peers, plus long-context and agentic suite results; "Although trained with only 12T tokens, Seed-OSS achieves excellent performance on several popular open benchmarks".

Verified 2026-09-20